Hallucinations aren’t random failures; they’re the price of ambiguity. Models fill in gaps with their best guesses. The fewer rules you provide, the more the prodigy improvises.
- Define invariants — Spell out non‑negotiables. These are the elements that must remain consistent across outputs:
- Spatial relations.
- Object lists.
- Attire, brand elements.
- Grammar/voice.
- Data schemas.
- Reduce the scope — Limit each instruction to a single unit of meaning:
- One shot, one scene, one action per instruction.
- Use time-codes for video.
- Function boundaries for code.
- Sections for prose.
- Itemize the steps — Break the process into discrete, reviewable stages:
- Shot list → storyboard → draft → final.
- Numbered steps, approvals between steps.
- Verify in loops — Inspect small outputs early and often:
- Stills, keyframes, code stubs.
- Reject precisely: what violated which rule.
- Don’t wait for the final product to catch errors.
- Enforce standards — Use negative constraints to prevent drift:
- Re‑run until acceptance criteria are met.
Creative cycle impact
Hallucinations don’t just distort outputs — they extend timelines. The creative process often runs longer than expected, not from lack of ideas, but from fixing, refining, or replacing outputs that miss the mark. Over multiple iterations, that friction compounds into extra time and cost.
Yet, not all surprises are unwelcome. In some cases — especially in code — hallucinations have produced features I hadn’t requested but ended up keeping because they were genuinely useful. These “happy accidents” are rare, but they remind us that while hallucinations usually add drag, they can occasionally expand the solution space in ways you wouldn’t have planned.
Hallucination examples
To illustrate these principles in action, here are three breakdowns:
cybercuriosity.tech 
Leave a comment